Professional Documents
Culture Documents
com/abstract=2003531
Quadratic Vote Buying
E. Glen Weyl
February 2012
Abstract
A group of individuals with access to transfers seeks to make a binary collective decision.
All known incentive compatible mechanisms are inecient (e.g. voting), subject to severe col-
lusion problems (viz. the Vickrey-Clarke Groves mechanism) or require the planner being fully
informed about the distribution of valuations (viz. the expected externality mechanism). I
propose a simple, budget-balanced mechanism with limited potential for collusion that achieves
eciency in large markets in any Bayesian equilibrium under (conditionally) symmetric inde-
pendent private values. Individuals purchase votes with the cost of a marginal vote being
linear in the number of votes purchased; thus the total cost of votes is quadratic in the number
purchased. The revenues earned from that individual are then refunded to other individuals.
Please do not cite this extremely preliminary draft without my permis-
sion. I have made it available online at the request of several colleagues,
but it is closer to a set of personal notes than a polished paper and the
senses in which this is true are discussed throughout the paper.
I am grateful to Eduardo Azevedo, Eric Budish, Drew Fudenberg, Scott Duke Kominers, Steven Levitt and
Stephen Morris for helpful comments.
Department of Economics, University of Chicago: 1126 E. 59th Street, Chicago, IL 60637: weyl@uchicago.edu,
http://www.glenweyl.com.
1
Electronic copy available at: http://ssrn.com/abstract=2003531
(D)emocracy is the worst form of government, except for all the others that have been
tried.
Sir Winston S. Churchill
As Churchills quote emphasizes, existing mechanisms for collective decision making are un-
satisfying. Voting, while robust in many ways (Hellwig, 2003), does not incorporate intensity of
preferences and thus is rarely ecient, in the sense that it decides in favor of the option maximizing
the sum of individuals willingness-to-pay.
1
Vickrey (1961), Clarke (1971) and Groves (1973) pro-
posed a mechanism which is ecient if individuals act unilaterally, but is highly sensitive to even
two individuals colluding and often requires destroying resources (Ausubel and Milgrom, 2005). All
other mechanisms I am aware of either require the administrator to have detailed information on
the distribution of valuations or have no general benets over voting or the Vickrey-Clarke-Groves
(VCG) scheme.
2
This paper proposes a novel mechanism for binary collective decision making with
transfers that is approximately ecient in any Bayesian equilibrium with a large number of indi-
viduals that have symmetric independent private values, is budget-balanced, puts strong limits on
the potential benets from collusion by a small group and has no information requirements for the
planner.
This Quadratic Vote Buying (QVB) mechanism is simple both to describe and implement. Each
individual i may purchase any continuous number of votes v
i
for her favored outcome at a cost v
2
i
.
Whichever alternative has more votes in its favor is selected and the expenditures by individual i
is return to all individuals other than i according to some pre-specied rule, such as even division.
Why is QVB ecient? The key argument is that, in any Bayesian equilibrium in a large market,
it forces individuals to pay their expected externality on others. Suppose that the distribution of
the sum of all other individuals votes is uniform. Then the second (unit of) vote purchases the
same marginal probability of changing the outcome as the rst vote does. However, it will change
the outcome in a case when, on average, there is one unit of vote in the opposite direction while
the rst vote will change the outcome, on average, when there is one half of a unit in the opposite
direction. Thus the marginal externality created by the second vote is, assuming proportionality
between aggregate votes and values, twice that of the rst vote. The marginal price of a vote should
thus be proportional to the number of votes already purchased, just as QVB does.
This argument relied on two assumptions. First, uniformity of the distribution of the sum of
others votes. Second, proportionality of the sum of others values to the sum of their votes. The rst
assumption is valid in a large market because any individual who is a small part of a large community
must view the distribution of the sum of others values (or votes) as approximately uniform over
1
In particular, it is only ecient (in large markets) when the threshold corresponds to the quantile of the distri-
bution corresponding to its mean (Ledyard and Palfrey, 1994). In small markets, eciency is even less common.
2
For example, the Expected Externality mechanism of dAspremont and Gerard-Varet (1979) and Arrow (1979)
requires the planner to know the exact distribution of values and the linear vote buying (storable votes) procedure
of Casella (2005) does not generally improve eciency.
2
the small range her values traverse. The second assumption is valid because the derivative of a
quadratic function is linear and thus individuals equate the number of votes purchased to half the
probability of their being pivotal times their valuation.
Why is QVB budget-balanced and resilient against collusion? All revenues are refunded to one
of the individuals, ensuring budget balance. Intuitively, collusion cannot achieve much: the best any
group of individuals can do is pool all their values and act as if they were a single individual, behaving
optimally for that individual. This leads to approximately to an aggregate outcome diering from
that without a collusion by at most multiplying the inuence of the group members by the group
size and saves them at most the payments they were making without collusion. So long as the group
is small compared to the total population, both of these eects are relatively small. This contrasts
sharply with VCG where any two individuals can costlessly assure their desired outcome.
This very preliminary draft is intended only to express the basic idea of and results on the
mechanism. The draft is thus divided into three sections, following the introduction. Section 1 lays
out the model I consider and formally describes the mechanism. Section 2 proves the main results
discussed above. Section 3 discusses the directions in which I hope to extend the results in a more
developed draft. Proofs are more like sketches and all details are omitted. Those interested in the
spirit of details prior to my release of an updated draft should see Carroll (2011) who develops
many of the tools needed to ll in additional details beyond the arguments included in the text. I
am also happy to correspond with parties about my thoughts on some of these.
1 Model
There are n individuals i = 1, . . . , n. A collective decision must be made as to whether to take
an action A or not. Let 1
A
represent the indicator function for whether A is undertaken. Each
individual receives a utility u
i
1
A
t
i
where u
i
is her value and t
i
is the (net) transfer she makes.
Individuals evaluate uncertain prospects as risk-neutral expected utility maximizers. Individuals
know their own values, but not the values of other individuals. Individuals values are drawn accord-
ing to independent and identical distributions (possibly contingent on some aggregate, commonly
known statistic) with continuously dierentiable probability distribution functions f
i
with support
on an open interval in R (possibly the whole line), as in Carroll (2011). I assume that the rst two
moments of f
i
exist and are and
2
respectively. I also assume E [|u
i
|] exists and denote it by |u|.
I dene eciency of a rule for choosing A as (2 1
A
1)
i
u
i
in a particular instance and its
expected eciency as the expectation of this, while ineciency in a particular instance is
_
1
i
u
i
>0
1
A
_
i
u
i
,
and its expectation is expected ineciency. Other denitions, with the exception of the proposed
3
mechanism dened below, are omitted but are standard.
Denition 1. The Quadratic Vote Buying (QVB) mechanism with shares {s
i
}
n
i=1
has
i
s
i
= 1
and allows each individual to choose how many (positive or negative) votes v
i
to purchase. It sets
1
A
= 1
i
v
i
>0
and t
i
= v
2
i
s
i
j=i
v
2
j
1s
j
.
Before getting to our results, note that QVB is trivially budget balanced as
i
t
i
=
i
v
2
i
s
i
j=i
v
2
j
1 s
j
=
i
v
2
i
j=i
s
j
1 s
i
v
2
i
= 0.
2 Results
A simple way to analyze the equilibrium properties of QVB is to consider it from the perspective
of any individual is optimization. Let G
i
be the cumulative distribution function of equilibrium
values of V
i
j=i
v
j
, with corresponding probability density function g
i
; by independence, all
values of other individuals are independent of is value and thus, so long as v
j
is measurable with
respect to u
j
alone (as I will show it is in equilibrium), V
i
will be independent of u
i
. If individual i
purchase votes v
i
the probability of the action occurring is 1G
i
(v
i
). Thus individual is ex-ante
expected payo is, up to the receipts from others vote purchases which she cannot inuence,
u
i
_
1 G
i
(v
i
)
v
2
i
.
The rst-order condition for maximization is then
u
i
g
i
(v
i
) 2v
i
= 0 v
i
=
u
i
g
i
(v
i
)
2
. (1)
Now suppose one could show that g
i
was independent of i and v
i
, at least for v
i
(v, v). Then
one would have that v
i
= ku
i
for some k > 0 and at any equilibrium one would have eciency as
the the outcome is determined by the sign of
i
v
i
which is the same as the sign of
i
u
i
. Even
if one could only show that g
i
changed by at most a factor one could bound the ineciency of
the outcome to within this factor. The rst subsection establishes a tight bound on this factor as
the size of the market grows. The second subsection uses these results to formalize and prove my
eciency claim. The third subsection establishes the sense in which the mechanism is robust to
collusion.
2.1 Uniformity
In this subsection, I assume, based on the logic above, that v
i
= ku
i
where k is a constant to be
determined in equilibrium. Then g
i
is just the sum of n 1 i.i.d. random variables with the
4
distribution ku
i
. Carroll (2011) extensively analyzes the properties of g
i
in this case. He makes
more precise than I will in this preliminary version the sense in which the normal approximation is a
valid approximation to the properties of g
i
. From here on out I simply assume this characterization
holds. In particular, I assume g
i
is a normal distribution with mean (n 1)k and variance
(n 1)k
2
2
.
Under this assumption, lets rst calculate g
0
g
i
(0) and the necessary value of k.
g
i
(0) =
e
[k(n1)]
2
2(n1)k
2
2
k
_
2(n 1)
=
e
(n1)
2
2
2
k
_
2(n 1)
.
If g
i
(v
i
) is approximately constant over the range of interest, it is approximately constant at the
value g
0
as 0 is included in the relevant range. From the logic above
k =
g
0
2
=
e
(n1)
2
2
2
2k
_
2(n 1)
k =
e
(n1)
2
4
2
2
4
_
2(n 1)
= g
0
=
4
2e
(n1)
2
4
2
4
_
(n 1)
.
I would like to argue that g
i
(v
i
) is very at over the range spanned by likely equilibrium values
of v
i
. I focus on the case when = 0 as, intuitively this is the most challenging case for eciency;
when = 0 the vote should be overwhelmingly in one direction or the other when n is large. In
reality, I have thus far been unable to fully analyze the case the = 0 case. The last subsection of
this section provides a very informal sketch of why I believe this case is stronger for eciency, but
trickier to establish.
When = 0, the normal density is at its maximum at 0 and thus its local slope is 0. Instead
one must consider its local second derivative which is
1
2s
3
. Thus the curvature of g
i
at this
point is
2
4
2
(n 1)
3
4
3
2
e
3(n1)
2
4
2
(2)
and thus g
i
(v
i
) diers by at most
4
v
2
i
4
2(n 1)
3
4
3
2
e
3(n1)
2
4
2
from g
0
.
2.2 Eciency
When = 0 even with aggregate information, no simple rule yields eciency for large n. Total
gains in this case are of order |
i
u
i
| which is, in expectation,
_
2n
n
g
2
0
8
u
2
i
.
Note that the joint distribution of n
i
u
i
and n
u
3
i
u
3
i
is, letting M
i
denote the ith moment
of the distribution of u (I assume moments through the 6th exist), for large n approximately
n
_
u
u
3
_
N
_
0
M
3
,
_
n
2
nM
4
nM
4
n(M
6
M
2
3
)
__
.
Let
n
be any function of n. Then ineciency is
nE
_
u
ng
0
4
u
3
<0
_
=
n
_
E
_
u
ng
0
4
u
3
<0
>
n
_
P
_
>
n
_
+ E
_
u
ng
0
4
u
3
<0
n
_
P
_
n
_
_
n
_
E
_
u
ng
0
4
u
3
<0
>
n
_
P
_
>
n
_
+ E
_
u
ng
0
4
u
3
<0
n
__
n
_
E
_
u
ng
0
4
u
3
<0
>
n
_
P
_
>
n
_
+
n
_
. (3)
Now I focus on the rst term.
E
_
u
ng
0
4
u
3
<0
>
n
_
P
_
>
n
_
= n
_
0
_ 4
ng
0
uh
_
u,
u
3
_
d
u
3
d
u + . . . (4)
where the second term is the corresponding expression for when
u > 0 and h is the density derived
from the joint normal distribution we discussed above, which I omit writing out explicitly to save
space. I focus on the rst term of the expression; the same logic can be used to bound the second
expression.
Conditional on
u,
u
3
N
_
M
4
u,
M
6
M
2
3
M
2
4
2
n
_
and thus the conditional probability that
u
3
4
ng
0
u is
_
_
4
ng
0
M
4
M
6
M
2
3
M
2
4
2
_
_
. For large n this is approximated by
_
n
u
_
where is a group
of constants as
n
g
0
is O
_
1
n
_
. By a well-known inequality
_
n
u
_
1
n
2n
u
e
2
n
3
u
2
2
. Whenever
6
u >
n
,
_
n
u
_
1
n
2nn
e
2
n
3
2
n
2
. Thus, for large n, we can bound the rst term from the
right hand side of Equation 4 by
n
n
2n
n
e
2
n
3
2
n
2
_
0
u
ne
2
d
u =
1
2n
n
e
2
n
3
2
n
2
_
1
2n
.
Letting
n
= n
3
2
+
for any > 0, we have exponential die-o of this term. The second term of
Equation (3) then becomes n
1
2
+
, which is clearly O
_
n
1
2
+
_
and yields the following quite strong
theorem
Theorem 1. If = 0 ineciency is O
_
n
1
2
+
_
, for all > 0 while total achievable eciency is
(
n).
Thus, in the most relevant case, QVB is highly ecient in large markets: while total attainable
eciency grows with
n, the ineciency of QVB shrinks at (nearly) the same rate. Thus the
relative ineciency of QVB shrinks at a (arbitrarily close to a)
1
n
rate.
2.3 Collusion
The eectiveness of collusion against QVB is tightly bounded: at most it can multiply the inuence
of a collusive group of size m by m while increasing their payments by the same factor or have
no inuence on the outcome while eliminating all payments. Similarly, one individual who is able
to pretend to be two is at most able to double her inuence and payment. Either way, if the
collusive group or the number of individuals one person can split herself into is a small fraction
of the population, there is little eect either on eciency or on the revenues dispersed to those
outside the collusive group. This contrasts sharply with VCG, where regardless of the size of the
population, any two individuals, or any one individual pretending to be two, may achieve their
desired outcome at zero cost.
To see this, suppose some subset M of all individuals join a coalition and coordinate their actions
to maximize their joint utility. Let G
M
denote the CDF of the distribution of the sum of all votes
outside the coalition. The total expected utility of the coalition is then
iM
u
i
_
1 G
M
_
iM
v
i
__
iM
t
i
.
Let x
M
iM
x
i
for x = u, v, t. In this subsection I assume s
i
1
n
and thus that
m1
n1
of the
revenue collected from the group is remitted back to it, where m |M|. The payo of the group is
given by
u
M
_
1 G
M
(v
M
)
_
1
m1
n 1
_
_
v
2
_
M
.
7
Notice that this expression depends only on v
M
and (v
2
)
M
and always negative only the second
assuming m = n. Thus an optimal allocation of votes among the individuals in the coalition always
requires minimizing (v
2
)
M
subject to a given level of v
M
. Because the square is convex, this always
requires setting the votes of all individual in the group to be the same. I can then rewrite the payo
as
u
M
_
1 G
M
(v
M
)
_
1
m1
n 1
_
v
2
M
m
.
The rst-order condition is then
u
M
g
M
(v
M
)
2(n m)
m(n 1)
v
M
= 0 v
M
=
m(n 1)
2(n m)
u
M
g
M
(v
M
) .
When n is large (both absolutely and relative to m) this simplies to v
M
= m
u
M
g
M
(v
M
)
2
. So long
as m is not too large, by the same arguments as in Subsection 2.1, g
M
is approximately invariant
to v
M
over its range and is approximately the same as f
i
over this range. Rather than repeat
arguments similar to the previous section to provide bounds, I simply skip to the limit and assume
g
M
is identically g
0
.
The optimal strategy for the coalition is approximately v
M
= m
u
M
g
0
2
. On the other hand, if the
individuals acted non-cooperatively their net votes would be (approximately) v
M
=
u
M
g
0
2
. Thus the
impact on the aggregate vote position of the collusive group is
(m1)u
M
g
0
2
.
On the other hand, if the collusive group operated independently they would each choose v
2
i
=
u
2
i
g
2
0
4
. Thus the total change in expenditures by the collusive group are
m
2
(u
M
)
2
g
2
0
4m
(u
2
)
M
g
2
0
4
=
g
2
0
4
_
m(u
M
)
2
_
u
2
_
M
.
To interpret this quantity, note that its positive part is determined by u
M
and its negative part
by (u
2
)
M
. Thus for a given u
M
the dierence is maximized when all group members have the same
value and thus
m(u
M
)
2
_
u
2
_
M
= m(u
M
)
2
m
_
u
M
m
_
2
= (u
M
)
2
_
m
1
m
_
=
m
2
1
m
(u
M
)
2
Thus, in this case, collusion actually raises payments by approximately m(u
M
)
2
. On the other
hand suppose that u
M
= 0. Then the expression is (u
2
)
M
, which is a lower bound on how much
payments may fall. However, note that in this case v
M
= v
M
= 0. This analysis is summarized in
the following theorem.
Theorem 2. For large n, absolutely and relative to m, a coalition M of size m acting optimally
against rather than unilaterally, and assuming all other players act as in equilibrium, never leads to
ineciency greater than (m1)
iM
u
i
iM
u
2
i
and thus expected reductions
8
in payments are bounded above by
mg
0(
2
+
2
)
4
. So long as m does not depend on n, all ineciency
is therefore O(1).
A similar analysis applies to de-mergers. Suppose one individual i can pretend to be c individ-
uals. By the same analysis as above, she will choose to have each of her c representatives report
(approximately)
u
i
g
0
2
and thus she will cause ineciency of at most (c 1)u
i
while increasing her
total expenditures.
Theorem 3. For large n, absolutely and relative to c, an individual i who can pretend to be c
individuals will never lead to ineciency greater than (c 1)|u| and thus will not cause expected
ineciency greater than (c 1)E[|u|]. The individuals payments will always increase. Again if m
is not a function of n, lost revenues are O(1).
This compares to VCG where if the value space is unbounded, a collusion of any two individuals
can cause arbitrarily large ineciency (ineciency up to the total gains relative to a random rule).
Such ineciency is clearly (